The pages AI assistants actually cite
Run a few dozen local questions through the assistants and a pattern shows up in the citations. It is not the prettiest sites that get read. It is not the ones with the best photography, the newest framework, or the most pages. What gets cited is remarkably plain: pages that answer a specific question in plain text, near the top, without making the reader work for it.
That is an annoying finding if you just paid for a redesign. It is a useful one if you would rather fix three pages than rebuild a website.
An assistant is not browsing your site
It is assembling an answer from text it can extract, and it has to decide quickly whether your page contains a fact relevant to the question. Almost everything that reads as professional to a human is invisible or useless in that process:
- Text inside images — hours, service areas, price lists, certifications baked into a graphic.
- Content that only appears after a script runs, or behind a tab, accordion or carousel.
- A service area expressed only as a map pin, with no city named in writing.
- A PDF price sheet linked from a page that itself says nothing.
- Three paragraphs of positioning before the page says what you do and where.
Brochure page versus answer page
| Trait | Brochure page | Answer page |
|---|---|---|
| Opening line | “Your trusted partner since 1998” | “We replace residential AC systems in Pembroke Pines.” |
| Service area | A map widget | A written list of cities |
| Pricing | “Call for a quote” | A range, with what moves it |
| Hours | An image in the footer | Text, plus what after-hours means |
| The actual question | Answered on a PDF | Answered in the first paragraph |
The right column looks blunter and converts better with humans too, for the same reason it gets cited: somebody arriving with a specific problem finds out in one glance whether you are the answer. Positioning language is not wrong, it is just in the wrong place — put it after the facts, not in front of them.
The four pages worth writing
- 01One page per city you actually serve.
Named in the heading, named in the text, with something true and specific about working there — permit quirks, typical housing stock, drive times. A template with the city swapped reads exactly like a template with the city swapped.
- 02One page per high-value service, in customer language.
Not “HVAC solutions.” What the customer types: replacing an AC unit, a system that stopped cooling, a maintenance plan. Answer the question the title poses in the first paragraph.
- 03A real pricing page.
A range with the factors that move it beats “call for a quote,” both for people and for anything reading your site to decide whether to name you. If you truly cannot publish numbers, publish the shape of the decision.
- 04One page of the questions you answer on the phone all day.
Do you work weekends. Are you licensed. Do you finance. How fast can you come out. Each one a heading, each answer under it in plain text, no accordion.
Then make the facts agree with each other
The same facts are probably published in six places — your site, your Google Business Profile, a handful of directories, an old listing from a business name you stopped using. When those disagree, you have handed every system reading about you a reason to prefer somebody it is more sure about.
- Name, address and phone, character for character, everywhere.
- Hours, including what after-hours actually means, in text.
- Service area as a list of places, matching what your city pages claim.
- Licenses and credentials written out, not only shown as badges.
- One canonical business name — not three variants across three listings.
None of this is clever and all of it is cheap. It is also the part most service businesses skip, which is why the pass that fixes it tends to move things within a quarter.
Start by finding out what is being cited about you today. Ask ten questions a customer would ask, note which sources come back, and read those pages the way a machine would: first two sentences, plain text only, no images. Whatever you cannot learn that way, your prospective customer's assistant cannot learn either.
We test a defined set of local questions across agreed AI surfaces, with repeated samples and the evidence behind every row — then hand you a prioritized list of fixes.
See how Byldr GEO works →